Abstract: With the arrival of the multi-cloud era, cloud-based intelligent operations can detect and handle cloud platform failures in advance to ensure their high availability. Because of the complexity of cloud systems, operational data shows various temporal dependency and inter-metric dependency in data locality and data globality, which brings great challenges to multi-dimensional time series anomaly detection. However, most of the existing anomaly detection methods for multi-dimensional time series learn feature representation from normal time series data and detect anomalies based on reconstruction error or prediction error. These methods cannot capture the local and global information dependency of multi-dimensional time series at the same time, resulting in poor anomaly detection effect. To solve the above problems, a fusion learning based unsupervised anomaly detection method for multi-dimensional time series is proposed. The local features and global features of multi-dimensional time series are modeled simultaneously to obtain more abundant time series reconstruction information, and anomalies are detected based on reconstruction errors. Specifically, by introducing the self-attention mechanism to the temporal convolutional network, the proposed model pays more attention to the global features of data while modeling the partial correlation of data. In addition, the information sharing mechanism is added between the temporal convolutional module and the self-attention module to realize the information fusion, so as to better reconstruct the normal mode of multi-dimensional time series. Our experimental results on multi-dimensional time series real-world datasets show that the proposed method improves F1 score by up to 0.0882 compared with the previous multi-dimensional time series anomaly detection.